Multi-dimensional monitoring system and fault prediction method of on-load tap-changer for converter transformer

Abstract

The invention discloses a multidimensional monitoring system of an on-load tap-changer for a converter transformer and a fault prediction method, the system comprises a motor, a change-over switch core and other parts, and a data acquisition module is provided with a current, vibration and oil chromatography acquisition unit and a vacuum tube life monitoring unit. The current sensor collects motor current, the vibration sensor collects vibration signals of the change-over switch, the oil chromatography collection unit analyzes an oil sample, and the vacuum tube life monitoring unit evaluates the residual life of the vacuum tube. And a data transmission and processing analysis module. The fault prediction method comprises the steps of data acquisition, preprocessing, multi-dimensional data fusion, fault prediction model construction, feature extraction and selection, fault prediction and evaluation and the like. And screening characteristic parameters and fusion data by adopting a plurality of technical means, and constructing a machine learning model to predict faults. According to the invention, multi-source data acquisition and multiple analysis methods are integrated, the equipment state can be comprehensively reflected, the monitoring reliability is improved, faults can be predicted in advance, and the equipment fault rate is reduced.

Description

Technical Field

[0001] The present invention relates to the field of power equipment monitoring and maintenance applications, and in particular to a multi-dimensional monitoring system and a fault prediction method for an on-load tap changer for a converter transformer. Background Art

[0002] Converter transformers are key components of HVDC transmission systems. Their on-load tapchangers (OLTCs) are responsible for adjusting voltage under load to maintain stable power system operation. However, due to the frequent operation of OLTCs, which are subject to multiple electrical, thermal, and mechanical stresses, they have a high failure rate. A OLTC failure can cause the converter transformer to shut down, impacting the reliability of the entire HVDC transmission system and causing significant economic losses.

[0003] Common on-load tap-changer monitoring methods are often limited, relying solely on periodic oil sampling or simple electrical parameter measurements, which fail to provide a comprehensive, real-time understanding of the operating status. Furthermore, most fault diagnosis methods are post-fault, requiring only post-fault investigations. This makes it difficult to provide early warnings and fails to meet the high reliability requirements of modern power systems. Therefore, a multi-dimensional monitoring system and fault prediction method for on-load tap-changers used in converter transformers was proposed. Summary of the Invention

[0004] The present invention provides the following technical solution: a multi-dimensional monitoring system for an on-load tap changer for a converter transformer, comprising: Motor, switch core, insulated transformer oil chamber, data acquisition module, data transmission module and data processing and analysis module, the data acquisition module includes a current sensor, a vibration sensor, an oil chromatography acquisition unit and a vacuum tube life monitoring unit; A current sensor is installed on the power supply line of the motor. The current sensor is used to collect the working current of the motor in real time. The vibration sensor is installed on the outside of the head cover of the switch core. The vibration sensor is used to collect vibration signals in different directions of the switch and measure the amplitude and frequency parameters of the switch vibration; An oil chromatogram collection unit is installed outside the insulating transformer oil chamber, and includes an oil sample extractor and a gas chromatograph analyzer. The oil chromatogram collection unit is used to extract oil samples from the insulating transformer oil chamber through the oil sample extractor and analyze the composition and content of dissolved gas in the oil through the gas chromatograph analyzer; The vacuum tube life monitoring unit is used to monitor the operating voltage, current, and arcing time of the switch, and evaluate the remaining life of the vacuum tube in combination with the aging model. The data transmission module is installed inside the data acquisition module. The data acquisition module uses a combination of wired and wireless communication to transmit the collected data to the data processing and analysis module. The data processing and analysis module integrates a data preprocessing unit and a multi-dimensional data fusion unit. The data processing and analysis module is used to filter, denoise and normalize the collected raw data, and conduct comprehensive analysis of the data on motor current, vibration, oil chromatography and vacuum tube life through data fusion algorithm.

[0005] The present invention provides a multi-dimensional fault prediction method for an on-load tap changer for a converter transformer, based on the multi-dimensional monitoring system for an on-load tap changer for a converter transformer, comprising the following steps: S1 data collection: The current sensor, vibration sensor, oil chromatography acquisition unit and vacuum tube life monitoring unit in the data acquisition module can obtain motor current, vibration signal, gas composition in oil and vacuum tube status data in real time; S2 data preprocessing: The data collected in step S1 is transmitted to the data processing and analysis module through the data transmission module, and then the collected data is filtered, denoised and normalized by the data preprocessing unit within the data processing and analysis module to eliminate outliers and interference; S3 multi-dimensional data fusion: The principal component analysis and correlation analysis methods are used to screen the characteristic parameters, and the data processed in step S2 are fused by combining DS evidence theory and Bayesian network algorithm; S4 fault prediction model construction: Build and train a fault prediction model based on historical monitoring data and machine learning algorithms, divide the model into training sets, validation sets, and test sets to optimize model parameters. When training the model, divide the historical data into training sets, validation sets, and test sets. S5 feature extraction and selection: First, characteristic parameters related to the fault are extracted from the multi-dimensional monitoring data after fusion in step S3. Then, principal component analysis and correlation analysis methods are used to screen and reduce the dimension of the characteristic parameters to remove redundant features. S6 Fault Prediction and Assessment: The operating parameters of the on-load tap changer monitored in real time are input into the fault prediction model constructed in step S4. The fault prediction model outputs the probability of fault occurrence and the type of fault based on the received parameters, and continuously optimizes the model accuracy by comparing with the actual operating conditions.

[0006] Preferably, the current sensor is a Hall effect sensor, which is installed between the neutral line and the live line of the motor power supply circuit, and is connected to the data acquisition module via a shielded cable.

[0007] Preferably, the vibration sensor adopts a magnetic piezoelectric acceleration sensor, and the vibration sensor is installed at a symmetrical position outside the switch core cover. The vibration sensor calculates the frequency domain energy distribution of the vibration signal in real time through the FFT algorithm.

[0008] Preferably, the vacuum tube life monitoring unit monitors the arcing current waveform through a Rogowski coil, collects vacuum chamber pressure data in real time in combination with a vacuum degree sensor, and establishes a contact material evaporation rate model based on the Arrhenius equation to dynamically evaluate the remaining life.

[0009] Preferably, the vibration signal collected by the vibration sensor in step S1 is filtered out by a bandpass filter to remove low-frequency noise below 20 Hz and high-frequency interference above 1 kHz, retaining the mechanical resonance frequency band signal of 100-500 Hz, and using the kurtosis index to detect the bearing fault characteristics.

[0010] Preferably, the data preprocessing unit in step S2 adopts a sliding window normalization method to perform mean normalization and variance normalization on the data within every 10 seconds, and the window sliding step is 1 second. The sensor zero drift and dimensional difference are dynamically eliminated by the sliding window normalization method. The data preprocessing unit has a built-in self-test program to regularly calibrate the sensor zero drift.

[0011] Preferably, the principal component analysis method in step S3 extracts parameters whose cumulative contribution rates of the first five principal components are greater than or equal to 90%, and screens characteristic parameters that are highly correlated with the fault through the Pearson correlation coefficient matrix, eliminates features with redundancy greater than 0.8, and finally retains 8-12 key features to input into the fusion model.

[0012] Preferably, the machine learning algorithm in step S4 adopts a long short-term memory network, the number of input layer nodes of the long short-term memory network is 1.5 times the number of feature parameters after fusion, the number of hidden layer neurons of the long short-term memory network is two-thirds of the input layer, and the Adam optimizer is used for optimization during model training.

[0013] Preferably, the characteristic parameters in step S5 include: characteristics of the current change rate and harmonic content extracted from the motor current data, characteristics of the vibration amplitude, frequency distribution and kurtosis extracted from the vibration data, characteristics of the content change trend and gas production rate of each gas component extracted from the oil chromatography data, and characteristics of the aging degree extracted from the vacuum tube life data.

[0014] In summary, compared with the prior art, the present invention provides a multi-dimensional monitoring system and fault prediction method for on-load tap-changers for converter transformers, which has the following beneficial effects: 1. This invention integrates a current sensor, a vibration sensor, an oil chromatography acquisition unit, and a vacuum tube life monitoring unit to synchronously acquire multi-dimensional parameters such as motor operating current, mechanical vibration, gas composition in oil, and vacuum tube status. It also uses principal component analysis and correlation analysis to screen key characteristic parameters, and combines DS evidence theory and Bayesian network algorithms to perform fusion analysis on multi-source heterogeneous data. This overcomes the limitations of traditional single monitoring methods and can comprehensively reflect the electrical, thermal, and mechanical coupling status of the on-load tap changer, avoiding the risk of misjudgment of a single parameter and improving monitoring reliability and accuracy. 2. By constructing a fault prediction model based on machine learning, the present invention can use historical data and real-time monitoring data to predict possible faults of on-load tap changers in advance, providing operation and maintenance personnel with sufficient time to carry out equipment maintenance and repairs, effectively avoiding the occurrence of faults, reducing equipment failure rates, improving the operating reliability of converter transformers and the stability of the power system, and reducing power outage time and economic losses caused by equipment failures. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a structural schematic diagram of the present invention.

[0016] Figure 2 This is a flow chart of the fault prediction model construction of the present invention.

[0017] Figure 3 It is a flow chart of the fault prediction method of the present invention.

[0018] Description of reference numerals: 1. Motor; 2. Switch core; 3. Insulated transformer oil chamber; 11. Current sensor; 21. Vibration sensor; 31. Oil chromatogram acquisition unit. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0020] See also Figure 1 The present invention provides a technical solution, a multi-dimensional monitoring system for on-load tap-changers for converter transformers, comprising: Motor 1, switch core 2, insulated transformer oil chamber 3, data acquisition module, data transmission module and data processing and analysis module. The data acquisition module includes current sensor 11, vibration sensor 21, oil chromatogram acquisition unit 31 and vacuum tube life monitoring unit. A current sensor 11 is installed on the power supply line of the motor 1. The current sensor 11 is used to collect the working current of the motor 1 in real time. The current sensor 11 is a Hall effect sensor. The current sensor 11 is installed between the neutral line and the live line of the power supply line of the motor 1. The current sensor 11 is connected to the data acquisition module through a shielded cable. A vibration sensor 21 is installed on the outside of the head cover of the switch core 2. The vibration sensor 21 is used to collect vibration signals in different directions of the switch and measure the amplitude and frequency parameters of the switch vibration. The vibration sensor 21 adopts a magnetic piezoelectric acceleration sensor. The vibration sensor 21 is installed in a symmetrical position on the outside of the head cover of the switch core 2. The vibration sensor 21 calculates the frequency domain energy distribution of the vibration signal in real time through the FFT algorithm; The oil chromatogram collection unit 31 is installed outside the insulating transformer oil chamber 3. The oil chromatogram collection unit 31 includes an oil sample extractor and a gas chromatograph analyzer. The oil chromatogram collection unit 31 is used to extract oil samples from the insulating transformer oil chamber 3 through the oil sample extractor and analyze the composition and content of dissolved gas in the oil through the gas chromatograph analyzer. The vacuum tube life monitoring unit is used to monitor the operating voltage, current, and arcing time of the switching switch, and evaluate the remaining life of the vacuum tube in combination with the aging model. The vacuum tube life monitoring unit monitors the arcing current waveform through the Rogowski coil, collects vacuum chamber pressure data in real time in combination with the vacuum sensor, and establishes a contact material evaporation rate model based on the Arrhenius equation to dynamically evaluate the remaining life. The data transmission module is installed inside the data acquisition module. The data acquisition module uses a combination of wired and wireless communication to transmit the collected data to the data processing and analysis module. The data processing and analysis module integrates a data preprocessing unit and a multi-dimensional data fusion unit. The data processing and analysis module is used to filter, denoise and normalize the collected raw data, and comprehensively analyze the data of motor 1 current, vibration, oil chromatography and vacuum tube life through data fusion algorithm.

[0021] See also Figure 3 The present invention provides a multi-dimensional fault prediction method for an on-load tap-changer for a converter transformer, based on the multi-dimensional monitoring system for an on-load tap-changer for a converter transformer, comprising the following steps: S1 data collection: The current sensor 11, vibration sensor 21, oil chromatography acquisition unit 31, and vacuum tube life monitoring unit in the data acquisition module are used to obtain motor current, vibration signals, gas composition in oil, and vacuum tube status data in real time. The vibration signal collected by the vibration sensor 21 is filtered through a bandpass filter to remove low-frequency noise below 20 Hz and high-frequency interference above 1 kHz, retaining the mechanical resonance frequency band signal of 100-500 Hz, and using the kurtosis index to detect bearing fault characteristics. S2 data preprocessing: The data collected in step S1 is transmitted to the inside of the data processing and analysis module through the data transmission module, and then the collected data is filtered, denoised and normalized by the data preprocessing unit inside the data processing and analysis module to eliminate outliers and interference. The data preprocessing unit adopts a sliding window normalization method to perform mean normalization and variance normalization on the data within every 10 seconds, and the window sliding step is 1 second. The sliding window normalization method is used to dynamically eliminate the sensor zero drift and dimensional difference. The data preprocessing unit has a built-in self-test program to regularly calibrate the sensor zero drift. The specific process of the above method is as follows; Data reception and transmission: The data collected in real time by the current sensor 11, vibration sensor 21, oil chromatography collection unit 31 and vacuum tube life monitoring unit in the data acquisition module are transmitted to the data processing and analysis module through the data transmission module. Data transmission adopts a combination of wired communication and wireless communication to ensure stable data transmission; Data pre-processing unit starts: After the data processing and analysis module receives the data, it starts the internal data pre-processing unit. The data pre-processing unit is responsible for filtering, denoising and normalizing the collected raw data to eliminate outliers and interference; Filtering and De-noising: The data preprocessing unit first filters the received data to remove high-frequency noise and low-frequency interference. The vibration signal is initially filtered using a bandpass filter during the data acquisition phase. Here, the data preprocessing unit further confirms or adjusts the filtering parameters to ensure signal quality. Next, the data preprocessing unit uses an appropriate denoising algorithm (such as wavelet denoising or median filtering) to further eliminate noise components. Sliding window normalization: The data preprocessing unit uses a sliding window normalization method to normalize the data. Specifically, for every 10 seconds of data, the data preprocessing unit calculates the mean and variance of the data in that time period. Then, the data preprocessing unit uses the calculated mean and variance to perform mean normalization and variance normalization on the data in that time period, so that the processed data has zero mean and unit variance. The window sliding step is 1 second. That is, after processing the data of each 10-second window, the window slides forward 1 second to continue processing the data of the next 10-second window. Dynamically eliminate zero drift and dimensional differences: Through sliding window normalization, the data preprocessing unit can dynamically eliminate sensor zero drift and dimensional differences. Zero drift refers to the phenomenon in which the output signal of a sensor deviates from zero due to environmental factors or aging during long-term operation. Dimensional differences refer to the fact that data collected by different sensors have different dimensions and numerical ranges. Sliding window normalization ensures that the data within each window undergoes the same normalization processing, thereby eliminating the impact of zero drift and dimensional differences on data analysis. A self-test program calibrates sensor zero drift: The data preprocessing unit includes a built-in self-test program that regularly calibrates sensor zero drift. This program compares the sensor output signal with a preset baseline value to detect zero drift. If zero drift is detected, the program automatically adjusts the sensor's calibration parameters to restore output signal accuracy. This calibration process can be performed when the system is idle to avoid affecting real-time monitoring data. Pre-processed data output: After filtering, denoising and normalization, the data is output by the data pre-processing unit to the multi-dimensional data fusion unit of the data processing and analysis module for subsequent data fusion and fault prediction analysis; S3 multi-dimensional data fusion: The principal component analysis and correlation analysis methods are used to screen the characteristic parameters, and the data processed in step S2 are fused by combining the DS evidence theory and the Bayesian network algorithm. The principal component analysis method extracts the parameters whose cumulative contribution rate of the first five principal components is greater than or equal to 90%, and the characteristic parameters highly correlated with the fault are screened by the Pearson correlation coefficient matrix. The features with redundancy greater than 0.8 are eliminated, and finally 8-12 key features are retained and input into the fusion model. The specific implementation process of the above method is as follows; Principal component analysis method to screen characteristic parameters: Perform principal component analysis on the data processed by S2 to check the contribution of each principal component to the overall data. Select the principal components with higher contribution rates one by one until the cumulative contribution rate of the first few principal components selected reaches or exceeds 90%. Here, select the parameters of the first five principal components that reach this contribution rate. These selected parameters will serve as the basis for further screening. Correlation analysis method to screen characteristic parameters: Build a tool for measuring correlation. For the data processed by S2, build a Pearson correlation coefficient matrix tool. Use the Pearson correlation coefficient matrix tool to find characteristic parameters that are highly correlated with the fault. Redundant feature elimination: Check the feature parameters obtained by principal component analysis and correlation analysis to see if there is redundancy between them. If the redundancy between a feature and other features is greater than 0.8, the feature will be eliminated. Determine key features: After eliminating redundant features, count the number of remaining feature parameters to ensure that the final number of remaining feature parameters is between 8 and 12; Data fusion: The 8-12 key feature data obtained from the previous screening are sorted and prepared for fusion operation. The DS evidence theory and Bayesian network algorithm are used to fuse these key feature data. The specific fusion process is as follows; Preparation for DS evidence theory fusion: For the 8-12 key feature data to be fused, a set containing all possible results (all possible situations related to the fault state of the converter transformer on-load tap-changer) is identified. This set is the identification framework. Based on the source and nature of the key feature data, such as the current-related features from current sensor 11 and the vibration-related features from vibration sensor 21, a basic probability value is determined for each key feature data, representing an initial estimate of the support level of the feature data for different fault situations. DS evidence theory fusion operation: According to the rules of DS evidence theory, the basic probability assignments assigned to each key feature data are combined. This step aims to comprehensively consider the information provided by each key feature data, fuse them together through combination rules, and obtain a comprehensive basic probability assignment result, which represents the degree of support for different fault conditions after fusion. According to the basic probability assignment results after fusion, a decision criterion is determined. For example, if the comprehensive support level of a certain fault condition exceeds a certain threshold, it is considered that this fault condition is the most likely to occur. This threshold can be set according to actual conditions and experience. It does not require precise numerical calculations, as long as it can distinguish different fault possibilities; Preparation for Bayesian network algorithm integration: A Bayesian network structure is constructed based on the causal relationship between key feature data and its relationship to faults. For example, if the current change rate may affect the life of a vacuum tube, and if vacuum tube life is related to faults in the on-load tap changer used in the converter transformer, this relationship is represented in the Bayesian network. For nodes in the Bayesian network (including those corresponding to key feature data and those corresponding to fault states), the prior probability is determined based on existing knowledge, experience, or previous data statistics. This is the probability of each node being in a certain state before considering other evidence. Bayesian network algorithm fusion operation: The results of the DS evidence theory fusion are used as evidence in the Bayesian network and propagated within the network according to the Bayesian network propagation rules. This means that the probabilities of each node in the network are adjusted based on the evidence to reflect the impact of the new information on the probability of failure of the entire system. After the evidence propagates, the posterior probability of each fault state is calculated. That is, after considering all key feature data (through DS evidence theory and Bayesian network propagation), the probability of each fault state occurring is determined. The final decision is determined based on the calculated posterior probability. See also Figure 2 , S4 fault prediction model construction: A fault prediction model is constructed and trained based on historical monitoring data and machine learning algorithms. The model parameters are optimized by dividing the model into training sets, validation sets, and test sets. When training the model, the historical data is divided into training sets, validation sets, and test sets. The machine learning algorithm uses a long short-term memory network. The number of input layer nodes of the long short-term memory network is 1.5 times the number of feature parameters after fusion, and the number of hidden layer neurons of the long short-term memory network is two-thirds of the input layer. The Adam optimizer is used for optimization during model training. The specific process is as follows: First, a database is established based on the collected data. The data collected when the on-load tap-changer is operating correctly and when the on-load tap-changer fails are collated and established into a database. A fault prediction model is then constructed and trained based on the established database and combined with a machine learning algorithm. S5 feature extraction and selection: First, characteristic parameters related to the fault are extracted from the multi-dimensional monitoring data after fusion in step S3. Then, principal component analysis and correlation analysis methods are used to screen and reduce the dimension of the characteristic parameters to remove redundant features. The above characteristic parameters include: the current change rate and harmonic content characteristics extracted from the motor 1 current data, the vibration amplitude, frequency distribution and kurtosis characteristics extracted from the vibration data, the content change trend and gas production rate characteristics of each gas component extracted from the oil chromatogram data, and the aging degree characteristics extracted from the vacuum tube life data; S6 Fault Prediction and Assessment: The operating parameters of the on-load tap changer monitored in real time are input into the fault prediction model constructed in step S4. The fault prediction model outputs the probability of fault occurrence and the type of fault based on the received parameters, and continuously optimizes the model accuracy by comparing with the actual operating conditions.

[0022] This solution integrates the current sensor 11, the vibration sensor 21, the oil chromatography acquisition unit 31, and the vacuum tube life monitoring unit to synchronously acquire multi-dimensional parameters such as the operating current of the motor 1, mechanical vibration, gas composition in the oil, and the vacuum tube status. Principal component analysis and correlation analysis are used to screen key characteristic parameters. The solution also combines the DS evidence theory and the Bayesian network algorithm to perform a fusion analysis of multi-source heterogeneous data. This solution overcomes the limitations of the single traditional monitoring method and can comprehensively reflect the electrical, thermal, and mechanical coupling status of the on-load tap changer, avoiding the risk of misjudgment of a single parameter and improving the reliability and accuracy of monitoring.

[0023] This solution builds a fault prediction model based on machine learning and can use historical data and real-time monitoring data to predict possible faults of on-load tap-changers in advance, providing operation and maintenance personnel with sufficient time to carry out equipment maintenance and repairs, effectively avoiding the occurrence of faults, reducing equipment failure rates, improving the operational reliability of converter transformers and the stability of the power system, and reducing power outage time and economic losses caused by equipment failures.

[0024] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0025] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A multi-dimensional monitoring system for on-load tap-changers for converter transformers, characterized by: include: A motor (1), a switching switch core (2), an insulating transformer oil chamber (3), a data acquisition module, a data transmission module, and a data processing and analysis module, wherein the data acquisition module includes a current sensor (11), a vibration sensor (21), an oil chromatogram acquisition unit (31), and a vacuum tube life monitoring unit; A current sensor (11) is installed on the power supply line of the motor (1), and the current sensor (11) is used to collect the working current of the motor (1) in real time. The vibration sensor (21) is installed on the outside of the head cover of the switch core (2), and the vibration sensor (21) is used to collect vibration signals of the switch in different directions and measure the amplitude and frequency parameters of the vibration of the switch; An oil chromatogram collection unit (31) is installed outside the insulating transformer oil chamber (3), and the oil chromatogram collection unit (31) includes an oil sample extractor and a gas chromatogram analyzer. The oil chromatogram collection unit (31) is used to extract oil samples from the insulating transformer oil chamber (3) through the oil sample extractor, and analyze the composition and content of dissolved gas in the oil through the gas chromatogram analyzer; The vacuum tube life monitoring unit is used to monitor the operating voltage, current, and arcing time of the switch, and evaluate the remaining life of the vacuum tube in combination with the aging model. The data transmission module is installed inside the data acquisition module. The data acquisition module uses a combination of wired and wireless communication to transmit the collected data to the data processing and analysis module. The data processing and analysis module integrates a data preprocessing unit and a multi-dimensional data fusion unit. The data processing and analysis module is used to filter, denoise and normalize the collected raw data, and conduct a comprehensive analysis of the motor (1) current, vibration, oil chromatography and vacuum tube life data through a data fusion algorithm.

2. The multi-dimensional monitoring system for on-load tap-changers for converter transformers according to claim 1, characterized in that: The current sensor (11) is a Hall effect sensor, and the current sensor (11) is installed between the neutral line and the live line of the power supply circuit of the motor (1). The current sensor (11) is connected to the data acquisition module via a shielded cable.

3. The multi-dimensional monitoring system for on-load tap-changers for converter transformers according to claim 1, characterized in that: The vibration sensor (21) adopts a magnetic piezoelectric acceleration sensor. The vibration sensor (21) is installed at a symmetrical position outside the head cover of the switching core (2). The vibration sensor (21) calculates the frequency domain energy distribution of the vibration signal in real time through an FFT algorithm.

4. The multi-dimensional monitoring system for on-load tap-changers for converter transformers according to claim 1, characterized in that: The vacuum tube life monitoring unit monitors the arc current waveform through a Rogowski coil, collects vacuum chamber pressure data in real time in combination with a vacuum degree sensor, and establishes a contact material evaporation rate model based on the Arrhenius equation to dynamically evaluate the remaining life.

5. A multi-dimensional fault prediction method for an on-load tap-changer for a converter transformer, based on the multi-dimensional monitoring system for an on-load tap-changer for a converter transformer according to any one of claims 1 to 4, characterized in that: The following steps are involved: S1 data collection: The motor current, vibration signal, gas composition in oil and vacuum tube status data are acquired in real time through the current sensor (11), vibration sensor (21), oil chromatography acquisition unit (31) and vacuum tube life monitoring unit in the data acquisition module; S2 data preprocessing: The data collected in step S1 is transmitted to the data processing and analysis module through the data transmission module, and then the collected data is filtered, denoised and normalized by the data preprocessing unit within the data processing and analysis module to eliminate outliers and interference; S3 multi-dimensional data fusion: The principal component analysis and correlation analysis methods are used to screen the characteristic parameters, and the data processed in step S2 are fused by combining DS evidence theory and Bayesian network algorithm; S4 fault prediction model construction: Build and train a fault prediction model based on historical monitoring data and machine learning algorithms, divide the model into training sets, validation sets, and test sets to optimize model parameters. When training the model, divide the historical data into training sets, validation sets, and test sets. S5 feature extraction and selection: First, characteristic parameters related to the fault are extracted from the multi-dimensional monitoring data after fusion in step S3. Then, principal component analysis and correlation analysis methods are used to screen and reduce the dimension of the characteristic parameters to remove redundant features. S6 Fault Prediction and Assessment: The operating parameters of the on-load tap changer monitored in real time are input into the fault prediction model constructed in step S4. The fault prediction model outputs the probability of fault occurrence and the type of fault based on the received parameters, and continuously optimizes the model accuracy by comparing with the actual operating conditions.

6. The multi-dimensional fault prediction method for on-load tap-changers for converter transformers according to claim 5, characterized in that: The vibration signal collected by the vibration sensor (21) in step S1 is filtered out by a bandpass filter to remove low-frequency noise below 20 Hz and high-frequency interference above 1 kHz, retaining the mechanical resonance frequency band signal of 100-500 Hz, and using the kurtosis index to detect the bearing fault characteristics.

7. The multi-dimensional fault prediction method for on-load tap-changers for converter transformers according to claim 5, characterized in that: The data preprocessing unit in step S2 uses a sliding window normalization method to perform mean normalization and variance normalization on the data within every 10 seconds, and the window sliding step is 1 second. The sliding window normalization method is used to dynamically eliminate sensor zero drift and dimensional difference. The data preprocessing unit has a built-in self-test program to regularly calibrate the sensor zero drift.

8. The multi-dimensional fault prediction method for on-load tap-changers for converter transformers according to claim 5, characterized in that: The principal component analysis method in step S3 extracts parameters whose cumulative contribution rates of the first five principal components are greater than or equal to 90%, and screens characteristic parameters highly correlated with the fault through the Pearson correlation coefficient matrix, eliminates features with redundancy greater than 0.8, and finally retains 8-12 key features to input into the fusion model.

9. The multi-dimensional fault prediction method for an on-load tap-changer for a converter transformer according to claim 5, characterized in that: The machine learning algorithm in step S4 adopts a long short-term memory network, the number of input layer nodes of the long short-term memory network is 1.5 times the number of feature parameters after fusion, the number of hidden layer neurons of the long short-term memory network is two-thirds of the input layer, and the Adam optimizer is used for optimization during model training.

10. The multi-dimensional fault prediction method for on-load tap-changers for converter transformers according to claim 5, characterized in that: The characteristic parameters in step S5 include: characteristics of the current change rate and harmonic content extracted from the current data of the motor (1), characteristics of the vibration amplitude, frequency distribution and kurtosis extracted from the vibration data, characteristics of the content change trend and gas production rate of each gas component extracted from the oil chromatogram data, and characteristics of the aging degree extracted from the vacuum tube life data.

Images